- Even the man who runs ChatGPT-maker OpenAI worries about the influence of AI on 2024's election.
- Sam Altman appeared before Congress on Tuesday to testify about the societal risks of AI.
Here's a not-so-distant future worth considering: When Joe Biden faces off against Donald Trump, Ron DeSantis, Nikki Haley, or another GOP candidate in the 2024 US presidential election, how sure can we be that the entire process won't be riddled with misinformation turbocharged by artificial intelligence?
Not very.
That much became clear when the chief executive of ChatGPT-creator OpenAI, Sam Altman, made his first appearance before Congress on Tuesday to discuss the potential risks of AI.
"This is a remarkable time to be working on artificial intelligence," Altman told Congress. "But as this technology advances, we understand that people are anxious about how it could change the way we live. We are too."
Gary Marcus, machine learning specialist and professor emeritus of psychology and neural science at New York University, was blunt during his own testimony about what some of those changes might look like.
"Fundamentally, these new systems are going to be destabilizing," he told lawmakers. "They can and will create persuasive lies at a scale humanity has never seen before. Outsiders will use them to affect our elections, insiders to manipulate our markets and our political systems. Democracy itself is threatened."
The devastation caused by social media in America's recent political history could look like child's play by comparison to AI.
Welcome to the era of personalized disinformation
There are already signs that AI could help influence undecided voters.
The large language models that underpin chatbots like ChatGPT can predict public opinion with remarkable accuracy when fed specific media diets, according to a pre-print research paper by researchers from Google, MIT, and Harvard cited by Sen. Josh Hawley on Tuesday.
In the context of an election, this could lead to situations where corporate, government, or foreign entities take these accurate predictions on public opinion and use them to "fine-tune strategies" that influence the way a voter acts, Hawley said.
Hawley said that even technology "as prosaic as Google search" can influence undecided voters trying to get information in the final days of an election, given what he described as the "enormous effect" the ranking of Google search articles.
"This of course is orders of magnitude far more powerful, far more significant," he added.
Altman acknowledged that "the more general ability of these models to manipulate, to persuade, to provide one-on-one interactive disinformation" is one of his areas of greatest concern.
Perhaps more worryingly, chatbots may have the power to change people's political beliefs entirely if they are manipulated to give people false information.
Gary Marcus cited a recent article published by the Wall Street Journal titled "Help! My Political Beliefs Were Altered by a Chatbot!", which details how people "may not even know they're being influenced."
And subcommittee chair Sen. Richard Blumenthal demonstrated at the start of the hearing just how easy it can be to lull voters into a false sense of security. He played an audio clip that sounded just like him, but was just an AI trained on his speeches.
Even Altman thinks AI will make humans stupid
For now, Altman said, humans understand that AI is in its infancy and are aware that bots like ChatGPT routinely make mistakes.
What he is less sure of is that users will continue to double-check ChatGPT's answers as its underlying model improves.
"I worry that as the models get better and better, the users can have less and less of their own discriminating thought process," Altman said. In other words, a powerful but imperfect AI will encourage lazy humans to outsource their own critical thinking — at truth and democracy's expense.
In fraught scenarios, such as a presidential election, the margin for error is zero.
Altman correctly (and self-interestedly) called during the session for AI to be regulated, including a suggestion that AI-generated content is clearly labeled.
This would only work if swift action is taken. Marcus rightly pointed out that "we acted too slowly with social media'' in previous elections. The same slowness just won't cut it in a world running to embrace ChatGPT.